Author: James M. Balter, Lando S. Bosma, Jorge Tapias Gomez, Nishant Nadkarni, Mert R Sanbuncu, William Paul Segars, Ergys D. Subashi, Neelam Tyagi, Harini Veeraraghavan 👨🔬
Affiliation: University of Michigan, The University of Texas MD Anderson Cancer Center, Department of Medical Physics, Memorial Sloan Kettering Cancer Center, Carl E. Ravin Advanced Imaging Laboratories and Center for Virtual Imaging Trials, Duke University Medical Center, Cornell University, University Medical Center Utrecht, Memorial Sloan Kettering Cancer Center 🌍
Purpose: Develop patient-specific virtual digital twin (VDT) cohorts modeling physically realistic spatio-temporal gastrointestinal (GI) organs (stomach and duodenum) digestive motion.
Methods: Patient-specific VDTs as 4D image sequences consisting of 21 phases modeling digestive GI motions were generated on 3D patient scans using published analytical models. A multi-step semi-automated pipeline was created: (a) AI-based organs auto-segmentation with visual verification; (b) organ-specific medial axis extraction; (d) non-uniform rational B-spline (NURBS) surface extraction; (e) multi-phase motion simulation on the NURBS using analytical models; (f) extraction of ground truth deformable vector fields (DVFGT); and (g) creation of 4D MRI/ 4D CT sequence simulating patient-specific motion. Eight datasets including T2w, T1w radial golden angle stack-of-stars, and contrast-enhanced CTs, were analyzed. Phase0 3D scan of each dataset was deformably aligned to 3D scan from phase with maximum displacement using 6 different deformable image registration (DIR) methods. DIR methods were evaluated using target registration error (TRE), Dice Similarity score (DSC), and the Hausdorff Distance at 95th percentile (HD95). Doses on the T2W MRIs were warped and accumulated using DVFGT and DVFDIR to calculate dose warping errors.
Results: A travelling wave equation Amplitude (A) =16mm, Wavelength (λ) =55mm, Wavespeed (c) =5mm/sec were applied to all datasets to generate a 4D image sequence of stomach & duodenum contraction phases along with the DVFs. Additionally, for two datasets an additional travelling wave equation A=16mm, λ =40mm, c=8mm/sec was applied to the large bowel. Average TRE for all datasets ranged from: 2.37mm (ProRseg) to 3.15mm (Demons). Average % Dose warping error ranged from 4.67% (EVO) to 5.9% (VoxelMorph).
Conclusion: Patient-specific digital twins with ground truth DVFs modeling a variety of realistic GI motion enable the creation of synthetic MRIs that can be used for validation of deformable image registrations and dose deformation, and eventually provide training dataset for deep learning models.